Machine learning based predictive model to enhance human reliability analysis for risk assessment of nuclear reactors

Abstract The aim of Human Reliability Analysis (HRA) is to evaluate the likelihood of the human operator performing erroneous actions in terms of Human Error Probability (HEP). Human operators are vital in safely operating complex systems, such as nuclear power plants (NPPs). In order to account for the role of the human, HRA experts have developed a variety of methods primarily targeting the quantitative dimension of human behaviour, with limited focus on the qualitative. The most recent shift has been towards Machine Learning (ML)-based methods for predicting HEPs. Prominent among these new methods are Logistic Regression (LR) and Bayesian Networks (BN). Amongst ML methods, Artificial Neural Networks (ANN) are also a good candidate model but have not been adequately explored in HEP evaluation. This article presents a comparative analysis of the above three approaches. The results indicate that ANN based models can replicate the qualitative HRA in their HEP predictions. Therefore, if qualitatively appropriate HRA data is available, then ANN-based models, utilizing operator performance data maybe a good alternative for HEP evaluation. The comparative scheme developed in this article can serve as the basis for future comparison amongst methods that employ an amalgamation of qualitative and quantitative methods in HRA.